Open source · Postgres · Supabase

Paste your schema.
Get realistic, perfectly-linked data.

Deterministic. Correct on the first run. Every foreign key resolves, every CHECK passes, every run with the same seed is byte-identical. No LLM in the data path.

pip install synth-scalecopy Try it in the browser ↓

01 · Try it

Playground

Paste Postgres DDL (CREATE TABLE, FKs, CHECKs, enums). Capped at 15 tables / 1,000 total rows / 50 KB. The CLI has no caps. Your DDL is parsed, never executed.

Rows 50 Seed

02 · Why

Why not just ask Claude for a seed script?

01

Reproducible

An LLM writes you a probably-correct script that changes every time you re-prompt. Synth-Scale is seeded end to end: same seed, same bytes, today, tomorrow, in CI. You can hash the output and assert on it.
02

Constraint-correct at scale

Tables are generated in FK dependency order, so every foreign key resolves, composite keys stay unique, and CHECKs are satisfied by construction, then an independent validator re-checks every constraint. 50,000 rows behave exactly like 50.
03

Coherent data

Not a random grid: updated_at ≥ created_at, orders newer than their users, a shipped order has a shipped_at, Karachi stays in Pakistan, and a few power users own most of the orders, like real production data.
04

Not a Faker replacement, a layer above it

Faker generates a realistic name, email, or date with zero concept of your schema. We use Faker under the hood for exactly that, then add what you'd otherwise hand-build yourself: parsing your DDL, resolving foreign keys in dependency order, and satisfying every CHECK and UNIQUE constraint by construction.

03 · See it break (or not)

What "just ask an LLM" typically looks like

This is illustrative, not a transcript from one specific model or vendor — it's the failure pattern that shows up constantly when a seed script is hand-written or generated without any schema awareness: a foreign key that doesn't actually exist yet. Same shop schema as the playground above.

Typical seed script · breaks

INSERT INTO categories (name, created_at) VALUES
  ('Electronics', NOW()),
  ('Home & Garden', NOW()),
  ('Sports', NOW());

INSERT INTO products
  (category_id, product_name, price, stock, created_at)
VALUES
  (1, 'Wireless Earbuds', 59.99, 200, NOW()),
  (2, 'Garden Hose', 22.50, 80, NOW()),
  (4, 'Yoga Mat', 18.00, 150, NOW());

ERROR:  insert or update on table "products"
  violates foreign key constraint
DETAIL:  Key (category_id)=(4) is not present
  in table "categories".

Synth-Scale, same schema · real output

INSERT INTO "products"
  ("id", "category_id", "product_name", "price", "stock", "created_at")
VALUES
  (1, 3, 'Wireless Earbuds Pro', 71.99, 127, '2026-01-01T00:12:40'),
  (2, 3, 'Running Shorts', 25.99, 996, '2026-01-01T00:37:27'),
  (3, 2, 'Single-Origin Coffee Beans', 36.99, 944, '2025-07-07T14:26:53'),
  (4, 2, 'Pocket Poetry Anthology', 11.99, 387, '2025-12-23T02:10:08'),
  (5, 2, 'Mountain Bike Helmet', 41.99, 80, '2025-11-13T22:52:40');

✓ Every category_id above references a category that was actually inserted. Not hand-picked — this is unedited output from synth-scale --ddl shop.sql --seed 42.

04 · CLI

The real tool is the CLI

This page is the demo. The CLI parses your migrations, introspects live Supabase/Postgres schemas, and loads 50k constraint-correct rows straight into your database (proof below ↓). No caps, fully offline.

$ pip install synth-scale

# see it before you seed it: pretty 10-row preview, zero writes
$ synth-scale --ddl schema.sql --rows 100 --preview

# full run: CSV / SQL inserts / direct DB load, deterministic seed
$ synth-scale --ddl schema.sql --rows 1000 --seed 42 --format sql --out ./seed

# point it at a live database instead of a .sql file
$ synth-scale --from-db --db-url postgres://... --rows 1000

Read the docs on GitHub →

05 · Proof

50,000 rows, verified against a real Postgres database

Not a toy schema: UUID primary keys, a composite primary key, a composite UNIQUE, a self-referencing foreign key, and CHECK constraints on dates, enums, and cross-column math — the kind of schema that breaks naive generators.

50,000rows · 10 tables

2.0sto generate

2.8sto load into Postgres

0constraint violations

Reproduce it yourself — same fixture, same seed, same command:

$ pip install synth-scale
$ synth-scale --ddl hard.sql --seed 42 --format sql \
  --rows organizations=50,users=2000,categories=30,products=3000,inventory=4000,coupons=500,subscriptions=3000,orders=12000,order_items=25000,audit_log=420

Fixture: tests/fixtures/hard.sql in the repo. Full methodology and raw numbers: datagen_pkg/docs/postgres.md. Want to see what the data itself looks like, not just the stats? Try the playground above ↑.

06 · FAQ

Questions people actually ask

Does it use an LLM to generate the data?

No. Every row comes from deterministic, seeded generation: Faker-backed value generation plus constraint-satisfying logic on top. No API calls, no network dependency, no run-to-run drift.

Which databases does it support?

Postgres and Supabase today. DDL parsing, constraint handling, and the "Connect to Supabase" introspection are all Postgres-specific. MySQL and SQLite aren't supported yet.

Is my schema or connection string stored anywhere?

No. Pasted DDL is parsed in memory and discarded once the response is sent. A "Connect to Supabase" connection string is used for a single request and is never logged or written to disk.

What's the difference between this playground and the CLI?

The playground is a capped browser demo: 15 tables, 1,000 rows, no direct writes to your database. The CLI has no caps and can load rows straight into your own database.

Can I use it in CI?

That's the main use case. pip install synth-scale, run it against your schema with a fixed seed, and assert on deterministic output.

Is it free?

Yes. The CLI is open source and free. This playground is a free demo of the same engine.

07 · Feedback

Would you actually use this?

Two questions, no email required. Tells us whether to keep building this or go do something else with our time.

08 · Contact

Send us a message

Questions, bug reports, feature requests. Goes straight to the maintainer.

09 · Stay in the loop

Get launch updates

PyPI release, Supabase introspection, CI data contracts. One email when it ships. Nothing else.

Paste your schema.Get realistic, perfectly-linked data.